How to Use Protein Model Fine-Tuning

This service learns whole-protein properties from examples you have measured. It starts with a pretrained protein model and adapts it to your dataset. For example, you can predict a numerical enzyme activity with Regression, soluble versus insoluble with Binary Classification, or a localization category with Multiclass Classification. Several targets can be learned together, including datasets where some measurements are missing.

For a first sequence-based experiment, use ESM2-650M with Adapter tuning and leave the other settings at their defaults. SaProt and ProstT5 are the options when you want to use structures. Larger models and more flexible settings are experiments to compare, rather than guarantees of better predictions.

1. Prepare your CSV

Export a spreadsheet as CSV, with one protein per row. Required input columns are ID (a unique protein name) and Sequence (its amino-acid sequence). Add a column for each measured property or class label. Column names and class labels are case-sensitive. Use the same measurement units and category spelling throughout a target column. Do not include FASTA headers in Sequence cells.

For example:

ID,Sequence,Activity,Soluble
protein_1,ACDEFGHIK,0.82,1
protein_2,ACDEFGHIKA,,0
protein_3,ACDEFGHIKAA,0.45,1

This illustrates the format; three example rows are not a sufficient training dataset. In Targets, add Activity as Regression and Soluble as Binary Classification. The blank Activity cell means that measurement is missing, not zero. Every training row must have at least one known target. Each target needs at least two known training labels and known labels in both validation and test data. Every classification category must appear in training. These are minimum validation requirements; reliable learning and evaluation usually need substantially more examples.

For binary targets, prefer 0 for negative and 1 for positive. If you use other labels, they are sorted as text and the second label is considered positive; for example, no comes before yes, so yes is positive. Integer codes identifying categories should use a classification task, not Regression.

2. Add structures and grouping information when needed

For SaProt or ProstT5, choose Structure or Mixed and upload structures through Input Structures. Add Structure and Chain columns to the CSV: a cell such as protein_1.pdb identifies an uploaded file, and A selects its protein chain. The chain must exactly match that row's Sequence. Use unique filenames and one structural model per file; resolve sequence mismatches or missing residues before submission. Structure mode requires a structure on every row. Mixed mode allows rows with no structure; leave both Structure and Chain blank on those rows. Sequence mode uses only amino-acid sequences and works with all models. The chosen mode also applies when you later predict new proteins.

Add an optional Group column to keep related examples together, such as variants from the same parent protein or measurements from the same protein family. Identical normalized sequences are also kept together automatically. This helps avoid overly optimistic performance caused by near-identical proteins appearing in training and evaluation data. Similarity Threshold optionally groups similar sequences before splitting; it reduces this risk but is not a guarantee about every cross-split pair.

3. Choose dataset splits and training settings

The default split is 80% training, 10% validation, and 10% test. Training proteins teach the model. Validation proteins measure progress, select the best saved model, and determine early stopping. Test proteins provide a final independent evaluation. Percentages must add up to 100. To choose assignments yourself, fill Split on every CSV row with train, validation, or test. Related groups must stay in one split; their sizes may prevent exact percentage matches. All three splits must contain proteins, and each target needs labels in all three.

Start with the default model, tuning method, pooling, and component sizes. Adapter learns small added components while keeping the pretrained protein encoder fixed. LoRA learns compact changes inside it; QLoRA learns similar changes using a compressed original model. Full Backbone changes the entire encoder, produces a larger saved model, and is available for encoders up to 1.3B parameters. ESM2-15B requires QLoRA. Shared adapters and prediction heads learn with every method; Task Adapter Width of 0 disables the additional target-specific adapters.

Inspect validation results when comparing settings. Training error alone can be misleading: a model can memorize training examples and perform poorly on new proteins. Avoid repeatedly selecting settings based on the same test scores, because that makes the test set less independent.

4. Interpret your results

summary.csv reports the model, target count, realized split sizes, and best epoch. training_history.csv shows training loss and the combined validation score over time; lower is better. The combined score averages regression MAE divided by that target's training-data standard deviation and one minus classification F1. This puts different target types on a more comparable scale. For binary targets, F1 evaluates the positive class; for multiclass targets, F1 averages class-level performance. The saved model is from the epoch with the lowest combined validation score, which may differ from the best epoch for an individual target. Test data never selects it.

metrics.csv reports validation and test performance separately for each target. For regression, MAE is the average absolute prediction error and RMSE gives larger errors more weight; both are in the target's original measurement units, and lower is better. When available, R Squared measures improvement over a constant prediction and can be negative; Spearman measures agreement in ranking, with values closer to 1 indicating better agreement. For classification, accuracy is the fraction of correct labels; balanced accuracy gives each represented class equal importance. Precision measures how often a predicted class is correct, recall measures how many examples of a class are found, and F1 balances the two. Binary AUROC and AUPRC summarize ranking performance across decision thresholds; higher is better, and AUPRC depends on the prevalence of positive examples. Some metrics are unavailable when there are too few examples or only one represented class.

held_out_predictions.csv pairs predictions with observed validation/test labels. Regression has one row per protein and target. Classification has one row per possible class and its probability; the predicted label is repeated on those rows. A high probability expresses the model's confidence, not a guarantee of correctness; probabilities have not been calibrated against a separate calibration dataset. Missing observed labels stay blank. splits.csv lists each protein's assigned split.

best_model.pt contains the learned weights and settings needed to reuse the best model, including target labels and numerical scaling. The original pretrained model is not bundled, so this download is not a standalone copy of the entire model. Full-backbone tuning includes the learned changes to the whole encoder and therefore produces a substantially larger file. This saved model supports prediction; resuming training is not supported.

5. Predict new proteins

Open Protein Model Inference and enter this successful training job's ID. Upload a new CSV with ID and Sequence, plus Structure/Chain references and uploaded structures if required by the saved input mode. You do not need to supply measured target values or repeat the training settings. Predictions are most useful for proteins and properties similar to those represented in your training data; check promising candidates experimentally.